Libraries

Before we start, let’s load a few libraries.

rm(list = ls())

options(warn = -1)

library(knitr)
library(ggplot2)
library(caret)
library(doParallel)

registerDoParallel(cores = (detectCores() - 1))

With our libraries loaded we can start loading our data.

Data Loading

Let’s read in our data.

data.2015 = read.csv("data/2015.csv")
data.2016 = read.csv("data/2016.csv")
data.2017 = read.csv("data/2017.csv")
data.2018 = read.csv("data/2018.csv")

Now, we will only deal with regular season events. So let’s remove the playoffs from our datasets.

get.regular.season = function(data) {
    subset(data, isPlayoffGame == 0)
}

season.2015 = get.regular.season(data.2015)
season.2016 = get.regular.season(data.2016)
season.2017 = get.regular.season(data.2017)
season.2018 = get.regular.season(data.2018)

Here is a table of all the columns we shall keep and what we shall rename them to.

get.helpful.data = function(data) {
    data.frame(x = data$xCordAdjusted,
               y = data$yCordAdjusted,
               angle = data$shotAngleAdjusted,
               dist = data$shotDistance,
               type = data$shotType,
               typeNum = as.numeric(data$shotType),
               goal = data$goal,
               team = data$teamCode)
}

# type:
# 1 -> empty
# 2 -> BACK
# 3 -> DEFL
# 4 -> SLAP
# 5 -> SNAP
# 6 -> TIP
# 7 -> WRAP
# 8 -> WRIST

analysis.2015 = get.helpful.data(season.2015)
analysis.2016 = get.helpful.data(season.2016)
analysis.2017 = get.helpful.data(season.2017)
analysis.2018 = get.helpful.data(season.2018)

Now, we can remove incomplete cases and create our machine learning model’s giant data set.

analysis.2015 = analysis.2015[complete.cases(analysis.2015),]
analysis.2016 = analysis.2016[complete.cases(analysis.2016),]
analysis.2017 = analysis.2017[complete.cases(analysis.2017),]
analysis.all = rbind(analysis.2017, rbind(analysis.2016, analysis.2015))
analysis.all = analysis.all[complete.cases(analysis.all),]
analysis.2018 = analysis.2018[complete.cases(analysis.2018),]

Here’s what analysis.2018 looks like:

analysis.2018

Now, we need a few functions to help us select certain subsets of data. We’ll define three functions: get.team.data, get.shooter.data, get.goalie.data.

get.team.data = function(data, code) {
    subset(data, team == code)
}

Calculating Statistics

We can calculate a few statistics, like goal (effective) percentage for a certain shot. Let’s write a function to do that right now.

calculate.goal.percentage = function(data) {
    goals = sum(data$goal == 1)
    total = nrow(data)
    goals / total
}

So, for example, Penguins’s goal percentage against slap shots would be calculated as follows:

penguins.2018 = get.team.data(analysis.2018, "PIT")
penguins.2018.slap = subset(penguins.2018, typeNum == 4)
penguins.2018.slap.eff = calculate.goal.percentage(penguins.2018.slap)

Their goal percentage is 0.0530035.

Creating the Models

Using the caret package, we can build machine learning models to help us determine which shot type is best.

Let’s get started with training our control.

control = trainControl(method = "repeatedcv", number = 5, repeats = 3)

Now we will train a few different types of models. Here is a list of the models we will train:

• Neural Network
• K Nearest Neighbors

model.nnet = train(goal ~ . -goal -team -type,
                   data = analysis.all,
                   method = "nnet",
                   trControl = control)

## # weights:  36
## initial  value 34868.630462 
## iter  10 value 21035.218950
## iter  20 value 19451.876327
## iter  30 value 18791.718659
## iter  40 value 18718.264585
## iter  50 value 18659.968779
## iter  60 value 18585.497607
## iter  70 value 18559.318557
## iter  80 value 18542.281988
## iter  90 value 18523.966040
## iter 100 value 18515.062866
## final  value 18515.062866 
## stopped after 100 iterations

model.knn = train(goal ~ . -goal -team -type,
                  data = analysis.all,
                  method = "knn",
                  trControl = control)

Now our models have been made.

Testing the Models

Let’s test our models on the 2018 data. Here’s what the testing data looks like:

analysis.2018

Now, let’s get our predictions:

nnet.prediction = predict(model.nnet, newdata = analysis.2018)
knn.prediction = predict(model.knn, newdata = analysis.2018)

nnet.prediction.data = data.frame(analysis.2018)
nnet.prediction.data$predict = nnet.prediction

knn.prediction.data = data.frame(analysis.2018)
knn.prediction.data$predict = knn.prediction

So, our Neural Network data looks like:

nnet.prediction.data

Our K-Nearest Neighbors data looks like:

knn.prediction.data

Visualizing the Predictions

Let us visualize how distance and the goal prediction are related.

make.knn.dist.plot = function(data, primary, secondary, team) {
    name = paste(team, "Predicted Goal Probability versus Distance (KNN)", sep = " ")
    plot = ggplot(data) +
        geom_smooth(aes(x = dist, y = predict), method = "auto",
                 fill = secondary, color = primary) +
        labs(title = name,
             x = "Distance from Net",
             y = "Probability of Scoring") +
        theme_minimal()
    plot
}

Here’s the same function, but for the neural network data.

make.nnet.dist.plot = function(data, primary, secondary, team) {
    name = paste(team, "Predicted Goal Probability versus Distance (NNet)", sep = " ")
    plot = ggplot(data) +
        geom_smooth(aes(x = dist, y = predict), method = "auto",
                 fill = secondary, color = primary) +
        labs(title = name,
             x = "Distance from Net",
             y = "Probability of Scoring") +
        theme_minimal()
    plot
}

Now let’s visualize the frequency of a shot.

make.type.freq.plot = function(data, primary, secondary, team) {
  name = paste(team, "Frequency per Shot Type", seq = " ")
  plot = ggplot(data) +
    geom_bar(aes(x = type), stat = "count",
                   fill = secondary, color = primary) +
    labs(title = name, x = "Shot Type", y = "Count") +
    theme_minimal()
  plot
}

The next graph is comparing type and goal prediction.

make.knn.type.plot = function(data, primary, secondary, team) {
  name = paste(team, "Predicted Goal versus Shot Type (KNN)", seq = " ")
  plot = ggplot(data) +
    geom_jitter(aes(x = type, y = predict), fill = primary, color = secondary) +
    labs(title = name, x = "Shot Type", y = "Probability of Scoring") +
    theme_minimal()
  plot
}

Here is the neural net flavor of the last function.

make.nnet.type.plot = function(data, primary, secondary, team) {
  name = paste(team, "Predicted Goal versus Shot Type (NNet)", seq = " ")
  plot = ggplot(data) +
    geom_jitter(aes(x = type, y = predict), fill = primary, color = secondary) +
    labs(title = name, x = "Shot Type", y = "Probability of Scoring") +
    theme_minimal()
  plot
}